Transfer Learning Based on CCA-TrAdaBoost for Insurance Fraud Detection
Phannana Aiemsuwan, Supawadee Srikamdee · 2025
This research presents a novel approach integrating transfer learning with feature mapping techniques to address limitations of samples and imbalanced classes in automobile and health insurance fraud detection. Our methodology uniquely integrates Canonical Correlation Analysis (CCA) for optimal feature mapping between domains with TrAdaBoost for effective knowledge transfer-a combination not previously explored in fraud detection literature. The framework leverages knowledge from credit card fraud domains with abundant samples (where transaction patterns share statistical similarities with insurance claims) to enhance insurance fraud detection performance. Experimental results demonstrate significant improvements in detection ability, with$\mathbf{F 1}$-scores increasing by up to$\mathbf{0. 2 3 9}$compared to traditional methods. Statistical analysis confirms the significance of these improvements ($\mathbf{p}<\mathbf{0. 0 5}$). The findings are particularly beneficial for small or new insurers with limited samples, enabling them to develop effective fraud detection systems that can reduce unnecessary expenses from fraudulent claims.